Potential and limitations of efficient machine-learning wind downscaling for energy-relevant applications in mountainous environments

Accurately representing complex spatio-temporal wind fields in mountainous terrain requires high-resolution atmospheric models, but these come with substantial computational cost. Although generally less accurate than physics-based models, machine learning-based wind downscaling offers computationally efficient alternatives for many energy-relevant applications; however, its performance depends on training data and local conditions, limiting its broad applicability.We present an enhanced version of the deep-learning-based near-surface wind downscaling model Devine (Le Toumelin et al., 2023), trained on controlled atmospheric simulations over synthetic topographies covering a wide range of slopes and terrain features. Wind-direction-dependent descriptive features facilitate deployment across different mountainous sites. We evaluate the model using high-resolution atmospheric simulations and ground-based observations in two mountainous regions with contrasting climates and topography, performing a spatio-temporal assessment of its strengths and limitations.The enhanced Devine model reproduces fine-scale wind patterns for terrain-induced flow as used in the training data, demonstrating transferability across mountainous sites. Its rapid generation of high-resolution wind fields enables applications such as wind resource assessment, atlas generation, climate impact studies, and short-term operational forecasts for wind farm operation. Overall, the evaluation shows how the enhanced Devine model can guide energy-related applications, indicating where it performs reliably and where caution is needed due to unrepresented wind regimes such as large-scale pressure-driven flow.LeToumelin, L., Gouttevin, I., Helbig, N., Galiez, C., Roux, M., and Karbou, F. (2023). Emulating the adaptation of wind fields to complex terrain with deep-learning. Artificial Intelligence for the Earth Systems, 2(1):1–39.